DATASET_CARD.md

# Dataset Card � csharp-dotnet-cpt-v0

## Intended use
Continued pretraining of Qwen2.5-Coder-1.5B on C#/.NET code. Named slices allow
training separate LoRA adapters per domain.

## Source
- **Primary:** `codeparrot/github-code` (public Hugging Face parquet dataset; per-file
  `license` label derived from each repository's LICENSE file via a go-license classifier).
- 179 of 1126 source shards processed.
- **Not used (gated, access not granted on this account):** `bigcode/the-stack`,
  `bigcode/the-stack-v2-train`, `bigcode/the-stack-smol`, `bigcode/the-stack-dedup`.

## License policy
Only files whose recorded license is in the allowed set were retained:
`0bsd`, `apache-2.0`, `bsd`, `bsd-2-clause`, `bsd-3-clause`, `cc0-1.0`, `isc`, `mit`, `mitre`, `ms-pl`, `unlicense`.

Excluded: GPL, AGPL, LGPL, MPL, unknown, no-license, proprietary.

### Tokens per license
- `mit`: 460,905 files, 430,391,064 tokens
- `apache-2.0`: 167,699 files, 204,609,979 tokens
- `bsd-3-clause`: 31,535 files, 36,707,726 tokens
- `bsd-2-clause`: 10,346 files, 15,022,311 tokens
- `unlicense`: 8,238 files, 6,533,139 tokens
- `cc0-1.0`: 1,881 files, 1,586,022 tokens
- `isc`: 822 files, 437,927 tokens

## Filtering
- Kept C# (`.cs`/`.csx`) and .NET build/config files (`.csproj`,`.sln`,`.props`,
  `.targets`,`.editorconfig`,`.config`, `global.json`, `nuget.config`, `appsettings.json`, ...).
- Removed: binaries, generated code (`<auto-generated>`, `.designer.cs`, `.g.cs`, `do not edit`),
  minified (median/max line length thresholds), vendored/large data blobs, secrets.
- Size window 40�300000 chars.
- **Exact dedup** by SHA-256 of normalized content; **near-dedup** by MinHash LSH
  (128 permutations, 16 bands, Jaccard >= 0.8).

### Dedup / rejection stats
```json
{
  "exact_duplicates_removed": 0,
  "near_duplicates_removed": 19828,
  "rejected_bad_extension": 17215737,
  "rejected_license": 263775,
  "rejected_generated_or_size": 113696,
  "rejected_secret": 949
}
```

## Slices
- `core_csharp_bcl_async_performance`: 424,888 files, 321,910,658 tokens
- `testing_debugging_repair_refactoring`: 79,215 files, 149,175,161 tokens
- `legacy_dotnet_migration_interop`: 68,059 files, 100,024,228 tokens
- `desktop_wpf_winui_maui`: 54,819 files, 71,724,689 tokens
- `efcore_sql_data`: 18,638 files, 16,996,348 tokens
- `architecture_di_auth_configuration`: 8,893 files, 10,461,997 tokens
- `distributed_dotnet_grpc_signalr_workers`: 9,208 files, 10,058,794 tokens
- `blazor`: 3,184 files, 5,529,264 tokens
- `aspnetcore_web_api`: 8,990 files, 5,056,231 tokens
- `nuget_msbuild_ci_tooling`: 5,532 files, 4,350,798 tokens

## Splits (repository-level)
A repository's files always map to one split via a deterministic hash of `repo_name`,
so no repository appears in more than one split.
{
  "test": {
    "files": 9711,
    "tokens": 9151159,
    "repos": 1061
  },
  "train": {
    "files": 662253,
    "tokens": 676849804,
    "repos": 66796
  },
  "validation": {
    "files": 9462,
    "tokens": 9287205,
    "repos": 1012
  }
}

## Secret scanning
{
  "files_quarantined": 949,
  "by_reason": {
    "secret": 949
  },
  "unresolved_findings_in_corpus": 0,
  "note": "Every retained file passed the secret scanner; matches were excluded (quarantined), so zero unresolved findings remain."
}

## Schema
```
text, repo_name, path, source_url, commit_or_blob_id, license,
source_dataset, content_sha256, quality_score, category_slice, token_count,
is_synthetic, collected_at
```

## Provenance
- `source_dataset`: `codeparrot/github-code`.
- `source_url`: HEAD-based GitHub blob URL (`https://github.com/<repo>/blob/HEAD/<path>`).
- `commit_or_blob_id`: `ghcode:<shard>:<sha12>` retrieval id.
- `content_sha256`: SHA-256 of normalized content.
- **Limitation:** codeparrot/github-code does NOT carry an upstream per-file commit/blob
  SHA; the recorded `commit_or_blob_id` is the dataset-internal retrieval id, not a GitHub
  commit hash. Enriching with exact upstream commits requires per-repo resolution (deferred).

## Storage
Zstandard (level 9) Parquet; data shards ~300 MB compressed.
Compressed total: 1.19 GB;
uncompressed: 7.10 GB.
Verified shards: 180/184.

## Reproducibility
Pipeline: `src/pipeline.py` (download->filter->dedup->classify->tokenize->shard->upload),
`src/finalize_splits_reports.py` (splits + reports). Deterministic split + MinHash seeds.